OpenADMET Models documentation

openadmet-demos

1.0 Curating CYP3A4 inhibition data from ChEMBL

Data tends to be scarce and sparse in the biochemistry world.

Due to the time consuming and costly nature of lab experiments, it is not always feasible to generate the tens to hundreds of thousands of data points potentially needed to train machine learning models (depending on the complexity of the model and the nature of the data).

We therefore want to be able to supplement our model training with data from external, publicly available sources, such as PubChem BioAssay, ChEMBL, etc.

To make these datasets amenable to training, they have to be acquired, processed, and cleaned. This protocol will walk you through the necessary bare minimum data transformation and cleaning so that we can train our CYP3A4 inhibition model.

CYP3A4 inhibition Data in ChEMBL

ChEMBL is one of the most widely used publicly available bioactivity databases. It aggregates experimental data from the literature, including assays measuring enzyme inhibition, receptor binding, and more. For CYP3A4, the relevant data is typically in the form of:

1. \(IC_{50}\) – the concentration of a compound required to inhibit 50% of CYP3A4 activity. Often expressed on a logarithmic scale as :

   \[pIC_{50} = -\log10(IC_{50} (M))\]

2. \(K_{i}\) – the inhibition constant, often derived from enzyme kinetics experiments.

3. Percent inhibition – sometimes assays report only the percentage of enzyme activity inhibited at a given concentration. These are often less useful than “Dose-response” measurements that assay the behaviour of an enzyme under a variety of substrate concentrations.

A whole host of other activity types are available, called a standard_type in ChEMBL terminology:

1. EC50 - Concentration of compound that produces 50% of maximal biological activity.

2. AC50 - Concentration of compound that achieves 50% of assay-defined activity.

3. XC50 - Concentration of compound that affects biological activity by 50%, but the exact type of effect is not specified.

4. \(K_{d}\) - Equilibrium constant, or the concentration where 50% of the target is bound to the ligand.

5. Potency - Measure of how much a compound is needed to produce a given effect.

For now we will limit ourselves to \(IC_{50}\) measurements as they most closely represents what we are trying to predict: CYP inhibition.

Most data in ChEMBL comes from published assays, each linked to a documented assay ID, a compound ID, and an experimental value. Importantly, the same compound can appear in multiple assays with slightly different values depending on assay conditions (e.g., substrate type, enzyme source, assay format).

How is data in ChEMBL organised

ChEMBL stores bioactivity data in a relational schema that is accessible via its web interface, REST API, or downloadable SQL databases. The most relevant tables/fields for us today are the target_id, the pchembl_value and the standard_type

Luckily for you we provide the tools to pull in data from ChEMBL easily in the OpenADMET Toolkit. We will use this tooling here to pull our activity data for CYP3A4 inhibition.

Disclaimer

There is no real one-size-fits-all approach to data curation given the variety of use cases. Some experimentation and iteration may be necessary.

Additionally this represents the MINIMUM level of curation needed to get a dataset like this fed into a machine learning model.

Additional processing steps dependent on the actual values (activity based curation) such as normalization, outlier detection etc. are likely required to get build the best models possible.

Using ChEMBL Target Curators

OpenADMET Toolkit provides a number of Curators to pull data from the ChEMBL database.

Some examples are listed here

PermissiveChEMBLTargetCurator: Curates activity data for a given ChEMBL target with a distinct ChEMBL target ID (e.g CHEMBL340 for CYP3A4)

SemiQuantChEMBLTargetCurator: Curates activity data for a given ChEMBL target including semi-quantitative data (<, <=, >=, =)

MICChEMBLCurator: Curates MIC e.g for bacterial hosts, fungal hosts, etc.

HepatotoxicityChEMBLCurator: Curates hepatotoxicity data.

etc.

This curator will now pull all activity measurements that have a pChEMBL value. Having a pChEMBL value gives a moderate amount of confidence that the datapoint is a relevant dose response assay and does not fall outside typical assay ranges. See the ChEMBL docs for more information on the curation process behind pChEMBL

We can use the PermissiveChEMBLTargetCurator to pull in our CYP3A4 data, which only requires that the user specify:

1. chembl_version - the version of ChEMBL database to pull from

2. target_id - the ChEMBL ID of the target protein

3. standard_type - the type of activity we’re interested in, e.g. IC50, EC50, etc.

from openadmet.toolkit.database.chembl import PermissiveChEMBLTargetCurator

chembl_version = 35 # Specify which version of the ChEMBL database you're accessing. 35 is the latest version as of 7/1/25
target = "CHEMBL340" # Specify the ChEMBL ID for CYP3A4. This needs to be looked up on the ChEMBL website.
standard_type = "IC50" # We want IC50s only as looking at CYP inhibition

# Instantiate the curator. This will automatically handle downloading the ChEMBL database
curator = PermissiveChEMBLTargetCurator(chembl_version=chembl_version,
                                        chembl_target_id=target,
                                        standard_type=standard_type,
                                        require_pchembl=True,
                                        require_units="uM")

# Now that the ChEMBL SQL database has been downloaded, we can load the activity data into a dataframe,
cyp3a4_data_raw = curator.get_activity_data()
cyp3a4_data_aggregated =  curator.aggregate_activity_data_by_compound()

This code can take a while to run as it needs to download the ChEMBL database.

Luckily OpenADMET has pre-curated data on a number of ADMET relevant targets already in our Data Catalogs repo. Check it out and explore the other ChEMBL data we have curated.

# Here's some data we prepared earlier!
import pandas as pd
cyp3a4_data_raw = pd.read_parquet("https://github.com/OpenADMET/data-catalogs/raw/refs/heads/main/catalogs/activities/ChEMBL_pChEMBL_IC50/ChEMBL35_IC50/ChEMBL_IC50_CYP3A4_CHEMBL340_raw.parquet")
cyp3a4_data_agg = pd.read_parquet("https://github.com/OpenADMET/data-catalogs/raw/refs/heads/main/catalogs/activities/ChEMBL_pChEMBL_IC50/ChEMBL35_IC50/ChEMBL_IC50_CYP3A4_CHEMBL340_aggregated.parquet")

cyp3a4_data_agg.head()

	OPENADMET_CANONICAL_SMILES	OPENADMET_INCHIKEY	assay_id_count	standard_value_mean	standard_value_median	standard_value_std	pchembl_value_mean	pchembl_value_median	pchembl_value_std
330	CC(=O)N1CCN(C2=CC=C(OCC3COC(CN4C=CN=C4)(C4=CC=...	XMAYWYJOQHXEEK-UHFFFAOYSA-N	69	2733.243333	57.00	13530.805498	7.088551	7.24	0.896828
334	CC(=O)N1CCN(C2=CC=C(OC[C@H]3CO[C@](CN4C=CN=C4)...	XMAYWYJOQHXEEK-OZXSUGGESA-N	51	491.523137	55.10	2977.186038	7.256471	7.26	0.564591
2093	CCOC1=CC=C(N2C([C@@H](C)N(CC3=CC=CN=C3)C(=O)CC...	WQTKNBPCJKRYPA-OAQYLSRUSA-N	16	15625.000000	14600.00	10297.281195	4.927500	4.87	0.365340
3035	COC1=CC=C([C@@H]2SC3=CC=CC=C3N(CCN(C)C)C(=O)[C...	HSUGRBWQSSZJOP-RTWAWAEBSA-N	11	29885.145455	12000.00	35187.813220	5.112727	4.92	1.060604
3711	ClC1=CC=C(COC(CN2C=CN=C2)C2=CC=C(Cl)C=C2Cl)C(C...	BYBLEWFAAKGYCD-UHFFFAOYSA-N	10	2205.634000	850.57	2423.189645	6.322000	6.07	1.191281

Using intake

We also make the same data available as curated by us using an intake catalog. Intake is a handy system to collect data in data lakes across a variety of formats and lends itself very nicely to scientific data and transformations thereof. Read more about intake at the docs.

import intake

cyp3a4_data_catalog = intake.open_catalog("https://github.com/OpenADMET/data-catalogs/raw/refs/heads/main/catalogs/activities/ChEMBL_pChEMBL_IC50/CATALOG_ChEMBL35_IC50.yaml")
cyp3a4_data_raw_intake = cyp3a4_data_catalog["CYP3A4_raw"]

# You can read from remote easily with the .read() method
materialised = cyp3a4_data_raw_intake.read()

materialised.head(5)

	assay_id	doc_id	standard_value	molregno	canonical_smiles	standard_inchi_key	tid	target_chembl_id	pchembl_value	compound_name	...	doc_journal	doc_doi	doc_title	doc_authors	doc_abstract	doc_patent_id	doc_pubmed_id	doc_chembl_release_id	OPENADMET_CANONICAL_SMILES	OPENADMET_INCHIKEY
0	44705	5045	15000.0	140586	Cn1cc(Cc2cn(CC(=O)N(CCN)Cc3ccc(-c4ccc(Cl)cc4)c...	FNUFKVJAONMVTK-UHFFFAOYSA-N	17045	CHEMBL340	4.82	None	...	Bioorg Med Chem Lett	10.1016/s0960-894x(02)00473-0	The discovery of SB-435495. A potent, orally a...	Blackie JA, Bloomer JC, Brown MJ, Cheng HY, El...	The introduction of a functionalised amido sub...	None	12182870	1	CN1C=C(CC2=CN(CC(=O)N(CCN)CC3=CC=C(C4=CC=C(Cl)...	FNUFKVJAONMVTK-UHFFFAOYSA-N
1	44705	5045	19000.0	141798	Cn1cc(Cc2cn(CC(=O)N(CCO)Cc3ccc(-c4ccc(Cl)cc4)c...	QBOWMQHTHPHCEL-UHFFFAOYSA-N	17045	CHEMBL340	4.72	None	...	Bioorg Med Chem Lett	10.1016/s0960-894x(02)00473-0	The discovery of SB-435495. A potent, orally a...	Blackie JA, Bloomer JC, Brown MJ, Cheng HY, El...	The introduction of a functionalised amido sub...	None	12182870	1	CN1C=C(CC2=CN(CC(=O)N(CCO)CC3=CC=C(C4=CC=C(Cl)...	QBOWMQHTHPHCEL-UHFFFAOYSA-N
2	44705	5045	8000.0	141296	CN(C)CCN(Cc1ccc(-c2ccc(Cl)cc2)cc1)C(=O)Cn1cc(C...	BFCVOTGOHBEFMZ-UHFFFAOYSA-N	17045	CHEMBL340	5.10	None	...	Bioorg Med Chem Lett	10.1016/s0960-894x(02)00473-0	The discovery of SB-435495. A potent, orally a...	Blackie JA, Bloomer JC, Brown MJ, Cheng HY, El...	The introduction of a functionalised amido sub...	None	12182870	1	CN(C)CCN(CC1=CC=C(C2=CC=C(Cl)C=C2)C=C1)C(=O)CN...	BFCVOTGOHBEFMZ-UHFFFAOYSA-N
3	44705	5045	10000.0	141018	CN1CCN(CCCN(Cc2ccc(-c3ccc(Cl)cc3)cc2)C(=O)Cn2c...	UYBIRMFTNJZIJL-UHFFFAOYSA-N	17045	CHEMBL340	5.00	None	...	Bioorg Med Chem Lett	10.1016/s0960-894x(02)00473-0	The discovery of SB-435495. A potent, orally a...	Blackie JA, Bloomer JC, Brown MJ, Cheng HY, El...	The introduction of a functionalised amido sub...	None	12182870	1	CN1CCN(CCCN(CC2=CC=C(C3=CC=C(Cl)C=C3)C=C2)C(=O...	UYBIRMFTNJZIJL-UHFFFAOYSA-N
4	44705	5045	5000.0	140588	CCN(CC)CCN(Cc1ccc(-c2ccc(Cl)cc2)cc1)C(=O)Cn1cc...	OQHPRXYGKQHOKG-UHFFFAOYSA-N	17045	CHEMBL340	5.30	None	...	Bioorg Med Chem Lett	10.1016/s0960-894x(02)00473-0	The discovery of SB-435495. A potent, orally a...	Blackie JA, Bloomer JC, Brown MJ, Cheng HY, El...	The introduction of a functionalised amido sub...	None	12182870	1	CCN(CC)CCN(CC1=CC=C(C2=CC=C(Cl)C=C2)C=C1)C(=O)...	OQHPRXYGKQHOKG-UHFFFAOYSA-N

5 rows × 31 columns

Exploring our dataset

Let’s have a quick look at the distribution of our inhibition IC50s: > They are mostly centered in the weak binding range (\(100\ \mu\text{M} - 1\ \mu\text{M}\)).

However, some compounds are quite potent inhibitors of CYP3A4 (\(< 1 \mu\text{M}\)) meaning that they will pose significant risk of causing drug-drug interactions.

import seaborn as sns
import matplotlib.pyplot as plt

# Plot settings
sns.set_theme(style="whitegrid")
plt.figure(figsize=(6, 4))

# Histogram with KDE overlay
ax = sns.histplot(
    data=cyp3a4_data_agg,
    x="pchembl_value_mean",
    bins=30,
    kde=True,
    color="teal",
    edgecolor="white",
    linewidth=1.2,
    alpha=0.85
)

# Titles and labels
ax.set_title("Distribution of CYP3A4 Inhibition Potencies (pChEMBL Values)", fontsize=18, weight="bold", pad=20)
ax.set_xlabel("Mean pChEMBL Value", fontsize=14, labelpad=10)
ax.set_ylabel("Compound Count", fontsize=14, labelpad=10)


# Make ticks bigger and easier to read
ax.tick_params(axis="both", labelsize=12)

# Subtle gridlines only on y-axis
ax.yaxis.grid(True, linestyle="--", alpha=0.6)
ax.xaxis.grid(False)

plt.tight_layout()
plt.show()

Lets have a look at some of the most potent inhibitors with mols2grid, a handy library for visualising chemical structures in a Jupyter Notebook.

import mols2grid

cyp3a4_data_agg_sorted = cyp3a4_data_agg.sort_values("pchembl_value_mean", ascending=False)
mols2grid.display(cyp3a4_data_agg_sorted, smiles_col="OPENADMET_CANONICAL_SMILES")

Final touches

Let’s add some finishing touches. It’s often nice to add a standard header to our dataset such that activity is standard across different datasets (standard_value provides this in ChEMBL).

At OpenADMET, we often use OPENADMET_LOGAC50 for this purpose.

Let’s also add a few columns describing our endpoint, OPENADMET_ACTIVITY_TYPE and target Target.

cyp3a4_data_agg_sorted["OPENADMET_ACTIVITY_TYPE"] = "IC50"
cyp3a4_data_agg_sorted["Target"] = "CYP3A4"
cyp3a4_data_agg_sorted["OPENADMET_LOGAC50"] = cyp3a4_data_agg_sorted["pchembl_value_mean"]

# save it to disk for use in next tutorial
cyp3a4_data_agg_sorted.to_csv("./processed_data/processed_CYP3A4_inhibition.csv")

Now let’s go on to train some models on our data!

End of 01_Curate_Data_ChEMBL.ipynb ~